Source-linked AI summary
Early- and in-season crop type mapping without current-year ground truth: generating labels from historical information via a topology-based approach
Chenxi Lin, Liheng Zhong, Xiao-Peng Song, Jinwei Dong, David B. Lobell, Zhenong Jin
TL;DR
Limited and delayed ground truth constrains timely crop mapping. The paper transfers historical topology in spectral feature space to generate target-year labels, rather than transferring decision boundaries. Across Iowa and Northeast China, the generated labels support accurate early- and in-season crop mapping.
Problem
Limited and delayed ground truth makes early- and in-season crop mapping difficult.
Method
The framework transfers crop-type topology in spectral feature space to generate target-year labels for subsequent classification.
Results
The approach generated high-quality labels across different crop types, sensors, years, and regions, with subsequent random-forest mapping achieving an Iowa overall accuracy of 0.8 early in the season.
Takeaways & Limitations
The framework supports transferring historical knowledge for timely crop mapping without relying on current-year ground truth.
Takeaways & Limitations
The study recognizes topological relationships between only two crops and identifies dependence on the percentage of target classes as limitations.
Abstract
from arXiv · showhide
Land cover classification in remote sensing is often faced with the challenge of limited ground truth. Incorporating historical information has the potential to significantly lower the expensive cost associated with collecting ground truth and, more importantly, enable early- and in-season mapping that is helpful to many pre-harvest decisions. In this study, we propose a new approach that can effectively transfer knowledge about the topology (i.e. relative position) of different crop types in the spectral feature space (e.g. the histogram of SWIR1 vs RDEG1 bands) to generate labels, thereby support crop classification in a different year. Importantly, our approach does not attempt to transfer classification decision boundaries that are susceptible to inter-annual variations of weather and management, but relies on the more robust and shift-invariant topology information. We tested this approach for mapping corn/soybeans in the US Midwest and paddy rice/corn/soybeans in Northeast China using Landsat-8 and Sentinel-2 data. Results show that our approach automatically generates high-quality labels for crops in the target year immediately after each image becomes available. Based on these generated labels from our approach, the subsequent crop type mapping using a random forest classifier reach the F1 score as high as 0.887 for corn as early as the silking stage and 0.851 for soybean as early as the flowering stage and the overall accuracy of 0.873 in Iowa. In Northeast China, F1 scores of paddy rice, corn and soybeans and the overall accuracy can exceed 0.85 two and half months ahead of harvest. Overall, these results highlight unique advantages of our approach in transferring historical knowledge and maximizing the timeliness of crop maps. Our approach supports a general paradigm shift towards learning transferrable and generalizable knowledge to facilitate land cover classification.
1. Introduction
Timely ground truth is a major bottleneck for early- and in-season crop mapping, while historical decision-boundary transfer can fail under changing spectral conditions. The study instead motivates transferring more stable topological relationships between crop classes to generate target-year labels.
- Motivation: Timely ground truth is a major hurdle for early- and in-season crop mapping because collecting it is costly and periodically updated maps may arrive months after the growing season.Ground-truth collection requires substantial time and labor, while products such as the CDL may not be released until January or February of the following year.
- Existing approaches: Historical decision-boundary approaches select spectral samples, train classifiers, and apply boundaries from prior years to target-year imagery.The workflow uses features such as Sentinel-2 RDEG1 and SWIR1, followed by classifier-based boundary construction and transfer.
- Existing approaches: Decision-boundary transfer is vulnerable because crop spectral features vary within and across years with weather, crop progress, and seasonal change.The problem persists even when phenological features are incorporated, limiting early- and in-season performance.
- Topology-based motivation: Soybean clusters remain northeast of corn clusters from early July across 2018 and 2019 despite changes in their absolute positions and shapes.This illustrates a time-invariant topological relationship in the RDEG1/SWIR1 feature space.
- Topology-based motivation: Stable topological relationships can be transferred from historical years to target years, potentially extending beyond corn and soybeans.The approach uses relative positions rather than directly transferring unstable classification boundaries to support target-year label generation.
2.1. Study area and data
The study evaluates crop mapping in Iowa and Northeast China using multi-year Landsat-8 and Sentinel-2 imagery, historical crop maps, and independent ground truth. The two regions were selected to test generalizability across differing climates, landscapes, and crop compositions.
- Iowa and the Songnen Plain in Northeast China were selected as study areas to test the approach across different climates, landscapes, and crop diversities.
- Iowa is dominated by corn and soybeans, while the Songnen Plain is characterized by corn, soybeans, and paddy rice.
- Sentinel-2 and Landsat-8 imagery were collected from June 1 through approximately September 30 across multiple years, using 5-day and 16-day composites, respectively.
- Historical CDL and cropland maps supplied training information, while field-collected or independent ground truth samples supported validation.
- The CDL and Northeast China cropland map are approximate predictions rather than perfect ground truth, but provide sufficient samples for learning topology knowledge.
2.2. Experiment overview
The experiments assess whether historical topology knowledge can generate target-year crop labels across years, sensors, regions, and crop types. The workflow trains a CNN-based model on historical heat maps, applies it at each target-year time step, and uses the resulting labels for crop mapping.
- The study evaluates topology knowledge transfer across different years, sensors, regions, and crop types through four experiments.
- Historical image patches and pseudo-ground-truth patches are converted into 2D histograms, which provide training inputs for recognizing crop topology.
- A CNN is trained to recognize topology relationships between crops and applied to each target-year time step to generate high-quality labels.
- The generated labels are then supplied to a classification algorithm for target-year crop type mapping.
- A random forest classifier is used so the experiments focus on topology-based label generation rather than advanced classifier performance.
2.3. Generating heat maps
The method represents image-patch pixels as density heat maps in a two-dimensional feature space and separates recognizable from unrecognizable crop topologies. It labels only the most confident crop regions and augments the heat map with background and coordinate channels.
- A heat map records the location and density of image pixels in a two-dimensional feature space derived from selected spectral features.
- Heat maps are categorized as type-I when they contain recognizable topology relationships and type-II when they do not.
- The method uses JM distance followed by visual inspection to categorize heat maps, although alternative measurements or visual interpretation alone can serve the same purpose.
- Only the most confident 50% of each crop distribution is labeled to extract pure and typical pixels for later classification.
- The 50% target removes corn–soybean overlap present in the 100% target, while type-II heat maps receive blank targets.
- Four input channels combine the heat map with historically uncultivated-pixel information and X/Y coordinates, helping constrain the potential trajectories of crop clusters.
2.4. Generating high-quality labels in the target year
A U-Net learns crop locations in the two-dimensional feature space from historical four-channel inputs and transfers that topology knowledge to target-year imagery. Recognized feature-space crops are converted back into spatial label patches for subsequent classification.
- A U-Net is trained as a supervised model to learn crop locations in the two-dimensional feature space from historical inputs and targets.
- The trained U-Net is applied to target-year four-channel inputs to transfer historical topology knowledge and recognize crops.
- Recognized crops are converted from feature-space outputs back into their spatial distributions within satellite-image patches.
- The resulting label patches provide crop labels for target-year classification, while blank patches indicate that crops were not recognized.
- Labels accumulate as imagery from later time steps becomes available, and the 50% training target may omit some soybean pixels without preventing further classification.
2.5. Crop type classification in the target year
The target-year classifier uses topology-generated crop labels, composited imagery, and a random forest to produce crop maps when labels are available. Evaluation compares generated labels with pseudo ground truth and classification results with field-collected ground truth.
- Twenty-day image composites from April 1 to the current date provide the classification imagery at each available time step.
- Twenty-day composites were selected because shorter periods were more cloud-contaminated, while 20 days matched 15-day performance with fewer features.
- Splitting each study region into nine grids mitigates spatial variability by prioritizing local labels and using neighboring-grid labels when necessary.
- A random forest maps crop types using topology-generated labels, all spectral bands, and NDVI, EVI, GCVI, and LSWI features.
- Generated-label accuracy is assessed against pseudo ground truth, whereas crop-map classification is assessed using field-collected ground truth.
- The study acknowledges that pseudo-ground-truth errors may propagate into generated-label evaluation, limiting its accuracy interpretation.
3. Result
Topology relationships across spectral feature spaces generate crop labels through the growing season, with Sentinel-2 generally providing strong agreements and timely classification results. The approach also achieves high accuracy for three crops in Northeast China, while sparser Landsat-8 observations constrain performance.
- Topology relationships: Corn pixels moved from below-left to directly below soybeans in RDEG1/SWIR1 space, whereas NIR/SWIR1 followed a more complex trajectory.
- Label generation: NIR/SWIR1 crop clusters were more variable and less concentrated than RDEG1/SWIR1 clusters, producing lower average label agreements.
- Iowa mapping: 0.929, 0.901, and 0.90 were the 2019 CDL corn F1, soybean F1, and OA values, while the RDEG1/SWIR1 experiment was 3.9%, 4.1%, and 2.7% lower.
- Sensor comparison: Landsat-8 reached maximum OA values of 0.847 in 2018 and 0.895 in 2019, below Sentinel-2 performance because of its 16-day revisit frequency.
- Northeast China mapping: F1 scores for paddy rice, corn, soybeans, and OA exceeded 0.85 by DOY 196, or July 15, in Northeast China.
4. Discussion
The discussion shows that topology-based labels consistently outperform transferred decision boundaries and postseason alternatives while enabling early crop mapping. It also identifies feature-space separability, sensor revisit frequency, crop abundance, crop-class scope, and historical reference-map availability as important conditions and limitations.
- Interpretation: The topology approach identifies transferable relative positions rather than inter-annually unstable classification decision boundaries.Its labels are generated from topologically stable features and then used for subsequent target-year classification.
- Benchmark comparison: 0.865 overall accuracy versus 0.778 in 2018 and 0.907 versus 0.827 in 2019 favored the proposed approach.These values compare the proposed labels with the benchmark approach.
- Further comparison: 0.813 and 0.873 overall accuracy from postseason full-season transfer remained below the proposed approach, while harmonic-regression transfer reached 0.713 and 0.853.The comparisons covered 2018 and 2019, respectively, and both postseason approaches used historical information without target-year labels.
- Limitations: The approach is constrained by two-crop experiments, sensitivity to target-class abundance, and reliance on pixel-wise historical reference products.Background can obscure minor crops in heat maps, while implementation may be limited where comparable historical products are unavailable.
5. Conclusions and outlook
The framework transfers historical crop knowledge through topological relationships in feature space and generates labels across crop types, sensors, years, and regions. Subsequent random-forest mapping achieves early crop-map availability, including Iowa results at least two months before harvest.
- The framework transfers historical knowledge through topological relationships among targets in feature space rather than relying solely on conventional classification transfer.
- The approach generated high-quality labels across different crop types, sensors, years, and regions for subsequent crop mapping.
- A simple random-forest classifier used the generated labels for subsequent crop type mapping.
- 0.8 overall accuracy was reached for Iowa corn/soybeans mapping as early as the silking or flowering stage, at least two months before harvest.
- Similar early-mapping results were obtained for paddy rice, corn, and soybeans in Northeast China.
- The framework can incorporate advanced deep-learning algorithms, including self-supervised and few-shot learning, in future extensions.